Crop residue cover (CRC), representing the fraction of residues retained on the soil surface after harvest, is vital for maintaining soil fertility, sustaining crop yields, and evaluating conservation tillage practices. Although hyperspectral satellite data show strong potential for CRC estimation, the capability of ZY1-02D imagery remains underexplored despite its high spectral resolution for distinguishing soil and crop residues. To fully leverage its capability, we developed a hybrid physics-informed deep transfer learning framework that integrates simulation-based pretraining with ZY1-02D hyperspectral satellite data for CRC estimation. A linear spectral mixture model (LSMM) was first employed to simulate soil-residue mixed spectra across varying CRC levels. A long short-term memory (LSTM) network was then pre-trained on the LSMM-simulated data and fine-tuned using in-situ CRC measurements and corresponding ZY1-02D spectra. For comparison, data-driven approaches including partial least squares regression (PLSR) and random forest (RF) were applied as benchmarks using both full-spectrum and feature-spectrum inputs. Results showed that the fine-tuned LSTM model, combined with standard normal variate preprocessing, achieved high accuracy ( R 2 = 0.84, RMSE = 0.12), outperforming the LSTM, RF, and PLSR models incorporating derivative features selected via the least absolute shrinkage and selection operator. Model interpretation using SHAP analysis and permutation-based feature importance further identified key wavelengths relevant to CRC retrieval. These findings demonstrate that integrating physics-informed spectral simulation with deep transfer learning significantly enhances CRC estimation from hyperspectral satellite imagery. This study provides a scalable approach for conservation tillage monitoring and advances the application of hyperspectral remote sensing in sustainable agricultural management. • Rapid and accurate estimation of crop residue cover is crucial to monitor conservation tillage and guide sustainable soil management. • A deep transfer learning framework is developed for estimation of CRC from ZY1-02D imagery. • Fine-tuned LSTM model outperforms RF and PLSR in CRC prediction with R 2 value of 0.84 and RMSE value of 0.12, respectively. • SHAP- and PFI-based interpretation identify key wavelengths related to crop residue–soil spectral differences.
Zhao et al. (Fri,) studied this question.